Topics Pioneers Insight Method Author
Back

Nvidia After Jensen Huang: AI, Chip Bans & Washington | Sharp Tech with Ben Thompson

·Sharp Tech

Rendering the long image…

Topics Mindset and Psychology of Tech Founders Physical Bottlenecks in AI Clusters Trillion-Dollar AI Capex and Bubble Risks US-China AI Rivalry and Geopolitical Controls One-Person Companies via End-to-End AI Autonomous Agents Reshaping Enterprise SaaS

Summary

  • The hosts frame Huang’s reported removal as a governance shock whose costs extend beyond Nvidia shareholders. Andrew Sharp puts the stock down about 15% over five days, despite a Thursday rebound, while Ben Thompson describes a $4.5–5 trillion company engineered around one decision-maker. Nvidia was “an F1 car that’s built around Jensen Huang”; replacing the driver raises questions that a conventional succession plan alone would not address.
  • Nvidia’s advantage rests on systems and supplier coordination as well as CUDA. Thompson credits the Mellanox acquisition with anticipating that AI would require interconnected chips, while Huang persuaded conservative Asian suppliers to invest ahead of demand. “It’s not a chip game anymore. It’s a systems game.” Losing the person who coordinated that system creates a different risk from losing a chip designer.
  • Startups and Neoclouds may lose their most powerful patron just as financing becomes harder. Huang supported open models and alternative infrastructure providers because Nvidia needed customers beyond hyperscalers capable of building competing chips. He also became the “bank for the bubble,” helping bridge AI investment from free cash flow through credit and equity toward eventual operating revenue; Thompson questions whether an interim CFO or future CEO will sustain that role.
  • Thompson argues that comprehensive chip restrictions manufacture the demand needed to displace America’s semiconductor ecosystem. Developers prefer Nvidia while it remains available; exclusion makes alternatives more necessary. His preferred mechanism is selling “N minus two, N minus three” generations: customers fund American R&D while retaining little incentive to build a competing ecosystem.
  • The hosts distinguish Huang’s business incentives from the allegations used to explain them. They question treating Alice Huang’s supposed status as a trained Chinese spy as established. Sharp says the disclosure sequence “feels pretty coordinated,” while Thompson’s defense of Huang’s commercial logic stops at legal compliance: “Laws are laws.”
  • The export-control disagreement ultimately turns on whether the long term still matters. Thompson sees a possible alignment between national-security hawks and people expecting decisive superintelligence within a few years. He rejects sacrificing durable industrial advantages to that forecast and questions how exclusive American superintelligence survives a disruption of Taiwan’s chip production: “What if you’re wrong?”
  • Thompson’s home workshop supplies a concrete example of agents creating demand and shortening physical production workflows. Claude assembled hundreds of small hardware purchases, hit a retailer’s 250-item cart limit, and helped design original parts for a $1,500 Bambu Lab H2C. A screen-and-keyboard workflow for applying inventory labels went from idea to operation in “20 minutes tops.”
  • Thompson revises his software thesis toward agents accessing systems of record rather than people opening bespoke apps. His inventory database remains useful even though he barely uses its interface: “I overstated the extent to which I’d want to create software.” The next constraint is execution access, illustrated by GPT-5.2 finally approving a Mac permission dialog through an external view of the machine.

Deep dive

1. The hosts see a shareholder loss inside a governance decision

  • Sharp opens with reporting attributed to The New York Times, citing five Nvidia sources, about an alleged affair with an executive and company resources used to conceal it. The Wall Street Journal subsequently reported that the board was preparing to oust Huang. The hosts say his removal became official Thursday morning; Sharp says he is wearing NVIDIA green “supporting the new era with Colette,” and later refers to Colette Kress being in charge.
  • Sharp’s puzzle: Nvidia shares were down about 15% over five days, despite a Thursday rebound, so why would shareholders benefit from removing Huang? Thompson applies the principal-agent problem to the board itself. Directors supposedly represent owners, but their incentives—including concern about personal liability—can diverge from what shareholders want. He believes most shareholders wanted Huang to remain.
  • Thompson nevertheless accepts that misconduct can justify removal on long-term fiduciary grounds. His categorical description of sleeping with a subordinate is: “You always get fired.” Sharp adds that directors could know things still outside the public record. Their objection concerns the consequences of the decision; neither presents Huang’s importance as an exemption from company policies.

2. CUDA emerged from bets that nearly killed the company

  • Thompson traces Nvidia’s history to the 1990s, an early failed graphics product, and a rescue from Sega in exchange for equity. The important technical progression was from graphics hardware executing predetermined drawing operations toward programmable shaders. Those small processors could eventually run programs themselves, opening possibilities beyond the original graphics workload.
  • The underlying opportunity was an “embarrassingly parallel problem”: divide a screen into smaller regions and perform many relatively simple calculations simultaneously. Thompson connects that architecture to AI’s computational demands. CUDA supplied the programming capability, but the initial motivation came from graphics rather than a fully formed prediction of today’s language-model market.
  • Adding programmability initially made Nvidia’s chips run hotter and less competitively. Thompson recalls ATI producing better graphics hardware while customers switched readily to whichever vendor performed best—an environment he compares with frontier-model switching today. CUDA was therefore not an immediately profitable moat; it was additional complexity that nearly sank the company before its broader usefulness became clear.
  • AlexNet’s image-recognition results were the wake-up call, although Thompson does not know whether Huang envisioned the broader opportunity before or afterward. Subsequent GTC demonstrations threw CUDA applications against the wall until ChatGPT supplied an overwhelming winner. Sharp remembers their 2021 interview as Huang still searching for how the technology would become widely useful.

3. The moat expanded from programming tools to complete systems

  • The AI boom also narrowed Nvidia’s story. Thompson remembers what he thinks was the 2024 GTC as “peak fame Jensen” and “least interesting Jensen”: product roadmaps displaced imaginative demonstrations. More consequentially, customers building on language models rather than directly on CUDA raised a question: how much protection would CUDA itself provide against alternative hardware?
  • Thompson credits the Mellanox purchase around 2019–2020 with anticipating the next layer: many chips functioning together as a larger computational system. Networking became essential to scaling. “It’s not a chip game anymore. It’s a systems game.” A competitor could therefore win an individual-chip benchmark without matching what Nvidia delivered across an interconnected installation.
  • His assessment of AMD is categorical on the distinction: strong hardware has repeatedly come with insufficient software and integration. He identifies TPUs as probably the largest credible training alternative at sufficient scale. He also says he believes Nvidia remains best on tokens per watt, which would favor it in an electricity-constrained world.
  • Supply was another moat. Huang secured memory and components while persuading conservative Asian suppliers that AI justified investment before demand was certain. Thompson’s sharpest example is Huang’s five or six annual keynotes: presentations directed at suppliers coordinating the tens of thousands of pieces in a rack. Huang called them “the most efficient way for me to communicate to you en masse.”

4. Neoclouds protected Nvidia against its biggest customers

  • Thompson’s value-chain framing: scarce chips let Nvidia capture exceptional margins, but abundance would strengthen buyers. Amazon, Google, Microsoft and frontier labs have the sophistication and resources to develop alternatives. A supplier dependent on a handful of such customers would eventually face pricing pressure, even if demand for computing continued growing.
  • His comparison is AMD’s progress against Intel in hyperscalers versus its limited penetration, in his telling, among enterprises and government buyers. A large cloud provider will invest in making another supplier’s processors work because supplier competition is valuable. Smaller organizations stick with the established platform rather than undertake that compatibility work themselves.
  • Nvidia therefore needed additional buyers committed to its ecosystem. Neoclouds, open models and startups expanded the market beyond customers with their own competing-chip ambitions. “NVIDIA did not want to become Intel.” Sharp emphasizes the resulting diversity; Thompson says startups losing the executive whose incentives supported that world is “a disaster.”
  • That patronage also involved power. Sharp points to CoreWeave receiving chips early, while Thompson questions whether allocations always maximized immediate economic returns. He explicitly does not know whether Amazon received fewer chips because of competing initiatives. His broader claim is that Huang made long-term ecosystem bets, potentially choosing winners to protect Nvidia’s future bargaining position.

5. Nvidia’s organization was designed around one generalist

  • Thompson’s personal encounters with Nvidia felt like dealing with a startup: improvised, fast-moving and charmingly ramshackle. Meta, by comparison, struck him as highly professionalized. Sharp observes that retaining startup characteristics at a roughly $4.5 trillion valuation sounds admirable. Thompson’s answer is that the same structure enabled Nvidia’s speed, annual release cadence and unconventional decisions.
  • Drawing on Tae Kim’s The NVIDIA Way, Thompson calls Nvidia “an F1 car built around Jensen Huang.” Sixty direct reports mean specialists feed decisions to the center. Their competence is not in question; the organization assigns the integrating role to Huang. Underneath its many demonstrations and applications, Thompson argues, Nvidia essentially sells one product—a GPU whose possibilities software expands.
  • Apple under Steve Jobs is his closest comparison. A functional organization hires specialists, then relies on its chief executive to integrate their work; a divisional organization distributes that responsibility among business units. Jobs’s famous reduction to four Mac products supported that structure. Thompson recalls criticizing Microsoft’s attempt to imitate it because Microsoft’s breadth required a different organizational design.

6. Succession requires an institution rather than another Huang

  • Thompson suspects Apple’s experience between 2008 and 2011 was essential preparation: Jobs’s illness allowed Tim Cook to run the company temporarily, followed by a period with Jobs back. The lesson was not to identify another Jobs. It was to create an organization that could survive without “an exoskeleton that depends on the guy in the middle.”
  • That comparison increases Thompson’s appreciation of Cook. Beyond supply-chain execution, he credits Cook with professionalizing Apple, including Apple University. He also recognizes the cost: writing down what makes a company distinctive begins a process of ossification. His analogy runs from walking through bicycles and cars to a bullet train—greater efficiency, progressively fewer directions available.
  • Nvidia instead built “the fastest freaking car on the planet,” retaining flexibility while neglecting governance and succession. Sharp compares it with a Red Bull tailored to Max Verstappen that teammates struggle to drive. Huang had recently discussed working another 20 years; Sharp cites talk of running Nvidia until 120. Neither sees evidence of Apple-like preparation for a sudden departure.

7. The hosts separate the spy narrative from the commercial incentives

  • The hosts describe a startling two-week collapse, but Thompson notes that Megaspeed had attracted New York Times and Bloomberg scrutiny a year earlier. Questions about a possible front and chips reaching China were longstanding. The allegation of a Chinese spy changed the story’s public reach; it did not introduce the underlying distribution issue.
  • Sharp corrects the name to Alice Huang. He describes the framing that she was recruited young and trained in the “dark arts”; Thompson says sexual blackmail is a tactic used by spy services, and Sharp says he does not think that was the case here. Sharp argues that someone moving billions of dollars of important equipment could plausibly become known to or be in contact with Chinese state authorities without that establishing she was a spy. “Everyone is a Chinese spy, and no one is a Chinese spy.”
  • Both reject the inference that the reported relationship explains Huang’s China policy. Sharp says Megaspeed represented about 1% of Nvidia revenue. Thompson says Huang’s longstanding commercial incentives to preserve Chinese customers and CUDA usage are the better explanation, and rejects explaining those incentives through Huang’s ethnicity.
  • On military access, Thompson accepts the uncomfortable trade-off: chips entering China could reach the PLA. He criticizes assurances that military use was impossible and explicitly owns that consequence of opposing comprehensive restrictions. Sharp agrees that civilian-versus-military buyers offer little reassurance and questions whether even a ban can reliably prevent determined military buyers from obtaining high-end hardware.

8. The first compliant workaround began a lasting Washington feud

  • Thompson traces the conflict to the initial Biden-era restrictions, which gated bandwidth rather than FLOPS. He disagreed with the policy but thought bandwidth was a sensible parameter for limiting large training systems. Nvidia produced a compliant, lower-bandwidth chip within a month. That chip was genuinely worse for scaling; Sharp’s suggestion that simply adding more chips restored equivalence gets rejected.
  • Washington, in Thompson’s interpretation, saw compliance with the letter rather than the spirit. Huang saw a businessman legally adapting a product to serve customers. “He was circumventing what you thought you did.” Nvidia’s limited Washington presence compounded the clash: officials came to regard Huang as a bad actor, while he regarded them as ignorant of the technology.
  • Huang subsequently gained access to Trump and helped reverse the H20 ban, the hosts say, while the earlier Biden restrictions remained. Sharp notes that China then would not buy H20s. Thompson sees continued appetite for Nvidia equipment through other channels as evidence that Chinese developers still preferred the hardware despite government efforts to move away.
  • Sharp says the disclosures “feel pretty coordinated,” intended to delegitimize Huang and his policy agenda. Thompson interprets the sequence—spy allegations, an employee affair, another affair, suggestions of worse revelations—as an endgame. Their account attributes significance to the timing; it remains their interpretation of how longstanding policy enemies converted personal misconduct into Huang’s removal.

9. Export bans give rival ecosystems the customers they need

  • Thompson’s objection to comprehensive restrictions centers on developer demand. A government can order an alternative to CUDA, fund it and announce its importance; it cannot manufacture an ecosystem by decree. “Ecosystems are pulled into existence by the developers who use them and want them and leverage them.” Those developers would generally prefer Nvidia if permitted to buy it.
  • His best analogy is Windows Phone. Microsoft spent what he guesses were billions trying to establish a platform, but Android and iPhones already met users’ needs. Money could not create sufficient pull. Similarly, an immature Chinese chip ecosystem struggles to recruit developers while established Nvidia products remain accessible. Exclusion changes that calculation from patriotic aspiration to operational necessity.
  • Thompson favors selling “N minus two, N minus three” generations. Older American hardware can still outperform a local startup’s offering, discouraging customers from investing time in that startup. Meanwhile, their purchases fund the supplier’s frontier R&D. He describes this as the longstanding technology-transfer approach: preserve a meaningful lead while getting disadvantaged customers to finance its continuation.
  • His long-run forecast is categorical: China will eventually develop manufacturing and CUDA equivalents, then export that ecosystem globally. He sees comprehensive bans accelerating this outcome at the expense of American semiconductor and equipment businesses. Sharp adds a present-day observation: despite repeated reports of Chinese manufacturing breakthroughs, demand for Nvidia—especially for training—has remained persistent.

10. The superintelligence timetable changes which costs policymakers count

  • Thompson suggests a coalition between national-security officials focused on China and people expecting imminent recursive self-improvement or superintelligence. Under the latter worldview, protecting semiconductor advantages over decades becomes irrelevant: “All that matters is the next two years.” He mocks that supposedly decisive two-year window for having persisted for roughly three years already.
  • Sharp says this was the governing logic behind the initial restrictions in 2022. Thompson places Huang on the opposite side: planning Nvidia’s next 50 years, supporting open models and resisting chip restrictions. He suspects Huang genuinely considered continued Chinese dependence on American technology good for the United States, while acknowledging how conveniently that belief aligned with Nvidia’s interests.
  • Thompson’s physical-world objection is Taiwan. A strategy assuming exclusive American superintelligence cannot ignore China’s ability to disrupt TSMC. His example is a reported issue involving high-speed-rail vibrations several kilometers from production facilities in Tainan. He says facilities that sensitive would not handle explosives well. “Where’s your ASI?” is his challenge to a strategy treating chip supply as secure.
  • The question is “What if you’re wrong?” about a short timetable that excuses sacrificing the long term. Sharp identifies resentment toward Huang across administrations, mentioning Marco Rubio and Scott Bessent under Trump; Thompson stresses the bipartisan character of the security establishment. He expects stronger restrictions and computer tracking after Huang’s exit, and regards the long-run industrial consequences as disastrous.

11. The AI investment bridge may lose its banker

  • Thompson places AI financing in a progression: free cash flow, then credit, then equity. “We’re well in the equity stage.” The necessary next step is sufficient revenue to repay obligations and restart investment from operating cash generation. That transition requires a bridge; infrastructure’s eventual usefulness does not itself pay today’s bills.
  • Huang supplied part of that bridge by supporting and backstopping companies—“being the bank for the bubble.” Thompson questions whether an interim CEO, particularly a CFO thinking about the numbers, will make comparable commitments, or whether a permanent successor can justify them. This is a question about sustaining risky ecosystem financing, not merely continuing Nvidia’s chip roadmap.
  • Neocloud credit conditions had already tightened, with concerns developing in the preceding months, Thompson says. Their weakness following the news therefore adds to an existing vulnerability. His railroad analogy matters: railroads were useful, yet financing structures could still collapse. Huang’s departure, in his assessment, significantly increases the probability of that kind of money crunch.

12. A sound commercial argument does not excuse ignoring the law

  • Sharp says he began as a skeptic of Thompson’s export-control argument and was eventually persuaded. He distinguishes Thompson’s approach—which still restricts semiconductor-manufacturing equipment—from Huang’s position. But agreement on long-term strategy does not resolve the immediate issue: Nvidia is an American company subject to American laws, whatever Huang thought of their wisdom.
  • Thompson explicitly marks his next claim as suspicion: Huang may have regarded the entire ban as illegitimate and treated apparent compliance as sufficient while chips continued reaching China. He suspects there is substantive wrongdoing behind the allegations. That differs from the earlier example of engineering an expressly compliant product; believing officials are wrong does not authorize disregarding their rules.
  • Sharp is frustrated precisely because he likes Huang’s optimism and recognizes his systemic importance. Thompson’s explanation is hubris: a fighter accustomed to winning, surrounded by weak governance, repeatedly ignored off-ramps while supplying personal ammunition to enemies. His tragic framing preserves the contradiction: the willingness to make enormous bets created Nvidia and also contributed to Huang’s downfall.

13. Agentic shopping turned a workshop ambition into actual purchases

  • Thompson’s practical AI example begins with a recurring frustration: driving about ten minutes to Ace Hardware and losing an hour obtaining a tiny quantity of bolts. He decided to build a hardware store in his basement. The $1,500 Bambu Lab H2C added another capability: “To vibe code an object” and then hold it was extraordinarily satisfying.
  • Claude proposed the fasteners, nuts and washers he might need. Thompson reduced an initial list costing several thousand dollars, then let it populate shopping carts with hundreds of small bags, typically ten- or 25-packs. One retailer capped carts at 250 items, requiring two orders. Thompson made the final purchase himself; the agent handled the tedious selection and assembly.
  • The distinction matters commercially: he says the agent “drove economic value” because he would not otherwise have undertaken all that shopping. It removed both the work of deciding what to stock and the friction of finding every variation online. Sharp calls the project excessive; Thompson accepts that it is also a hobby.

14. Original physical parts became another conversational workflow

  • Thompson ordered roughly 12 shelves with eight bins each, then discovered that the bins sat loose rather than sliding securely. His solution was a printed rail system: four interlocking pieces for each long shelf. He measured dimensions and discussed constraints with Claude, including side-to-side movement, front-to-back movement, shelf columns and a rear crossbar.
  • Sharp checks whether Claude merely retrieved an existing CAD model. Thompson says these were original designs, as were a keyboard tray and other objects Sharp had seen. Claude generated the model, handled slicing into printable layers and sent the work to the machine. “I don’t use 3D software. I don’t use Bambu Lab software.”
  • The workflow also included observing the printer’s camera, sending progress pictures and detecting when Thompson removed a completed item so the next print could begin. For him, the important change is the distance from an idea to execution. Physical fabrication now joins coding as something he can direct through conversation without first mastering another specialized interface.

15. The inventory project becomes a system of record with disposable interfaces

  • Because Claude had selected the hardware and planned its storage, it could populate Thompson’s inventory database before anything was labeled. The database records where objects belong rather than maintaining exact counts. Claude then obtained the bins’ label template and generated labels and QR codes from the existing records, reversing the earlier process of photographing and entering items individually.
  • Applying around 1,000 labels created a new bottleneck. His assistant Josh initially consulted a Telegram bot on his phone to learn each destination. Thompson saw the friction and had Claude put a visual guide on a networked screen: the relevant shelf, bin and cup appeared alongside the current label.
  • An old Bluetooth keyboard became a single “next” button: any key advanced the process. The printer produced a label, the screen showed its destination, Josh applied it, and a keypress requested the next one. Thompson puts the time from recognizing the problem to having that workflow running at “20 minutes tops.”
  • This leads to a revision of his earlier thesis about personal software. He barely opens the inventory app; he asks an agent that accesses its database. “I overstated the extent to which I’d want to create software.” The lasting need may be a “system of record,” with interfaces generated or bypassed according to the immediate task.

16. Access and permissions determine how much an agent can finish

  • Thompson still prefers ChatGPT for organized programming projects, including Gecko, while Claude better handles his overlapping conversations and household work. He praises “Opus 5.5, the GOAT” for speed and capacity, then describes GPT-5.2 successfully clicking a Mac permission dialog. His preferences are task-specific rather than a claim that one model wins every category.
  • A KVM gave the agent an external view of the Mac Mini, including permission dialogs that ordinary programs on the machine could not see. Previously, it identified the obstruction and messaged Thompson to approve it manually. This week, it finally acted on his standing instruction that this was its own machine: “It’s my Mini, so I clicked okay.”
  • The breakthrough mattered because agent-written software can otherwise stall on routine permissions. Thompson thinks it occurred while creating the keyboard remapping app. He also describes Claude refusing household-system access until Astra “broke in,” after which he handed Claude the key. His broader condition remains that agents need the right access and controls; capability alone does not complete the workflow.
  • Local inference is consequently becoming something Thompson feels unable to live without. He describes “Apple Harbor” as an inference-server mode, assumes it would avoid operating-system overhead and permission dialogs, and considers backup Studios. Apple’s software stack still trails CUDA in his assessment. Dependence on available open models leaves a final appeal: “Meta, please do some more open source.”